hazeremoval_singleimage_luminance
TRANSCRIPT
-
7/28/2019 HazeREmoval_SingleImage_Luminance
1/4
Automatic Image Haze Removal Based on
Luminance Component
Fan Guo Zixing Cai Bin Xie Jin TangSchool of Information Science and Engineering, Central South University,
Changsha, China
AbstractIn this paper, we propose a simple but effective method
for visibility restoration from a single image. The main advantage
of the proposed algorithm is no user interaction is needed, this
allows our algorithm to be applied for practical applications,
such as surveillance, intelligent vehicle, etc. Another advantage
compared with others is its speed, since the cost of obtaining
transmission map is greatly cut down by using Retinex algorithm
on luminance component. A comparative study and quantitative
evaluation is proposed with the main present-day methods whichdemonstrate that similar or better quality results are obtained.
Keywords-haze removal; YCbCr; luminance component;
Rentinex; transmission map
I. INTRODUCTIONHaze removal is highly desired in computer vision
application, removing haze can significantly increase thevisibility of the scene. Since the scene visibility is decreasedand the features, such as the contrast and color of the imageobjects are attenuated under haze condition, so we need toeliminate the haze effect of the scene objects.
Recently, the methods of image haze removal can beclassified into two major types: haze image enhancement basedon image processing and image restoration based on physicalmodel. The image enhancement method can effectivelyincrease the contrast and improve the visual effects, but maycause some lose to the stressed information. The classicmethods of this kind are histogram equalization, homomorphicfilter [1], wavelet transform [2], Retinex algorithm [3],luminance and contrast transform and so on. While the imagerestoration is researching on the degradation process of hazeimage, establishing the degradation model to get theundisturbed original image or its best estimator, so the qualityof the haze image can be improved. Compared with the imageenhancement method, this kind of method can get naturalresult, and assure no information will be lost to obtain the idealrestoration effect.
The degradation model that is widely used in imagerestoration method is the haze image model proposed byMcCartney [4], and a few approaches have been proposed.Very recently and for the first time in [5, 6, 7], threeapproaches were proposed which work from a single imagewith a stronger prior or assumption. Fattal [5] estimates thealbedo of the scene and then infers the medium transmission.This approach cannot well handle heavy haze images and mayfail in the cases that the assumption is broken. Tan [6] removes
the haze by maximizing the local contrast of the restored image,but his approach may not always achieve equally good resultson very saturated scenes. In comparison, Hes approach [7] isthe most attracting one. This approach based on the darkchannel prior. Using this prior, the thickness of the haze can bedirectly estimated and a high quality haze-free image can beobtained. However, the disadvantage of these three algorithmsis a processing time of 35 seconds, 5 to 7 minutes and of 10 to
20 seconds on a 600*400 image, respectively.
Here, we propose an automated dehazing approach which ismuch faster compared to [5, 6, 7]. This approach is based onthe observation that when the haze image transforms fromRGB to YCbCr color space, and uses Multi-Scale Retinex(MSR) algorithm to the luminance component, then doessubtraction operation and median filter on the map, thetransmission map whose function is similar to Hes approach isobtained. It is much faster since the cost of obtainingtransmission map is cut down, and it is able to achieve equallyand sometimes even better results.
II. BACKGROUNDA. Haze Image Model and Dark Channel Prior
Haze image model (also called image degradation model),proposed by McCartney [4] in 1975, the model consists ofdirect attenuation model and air light model. Direct attenuationmodel describes the scene radiance and its decay in themedium, while air light results from previously scattered lightand leads to the shift of the scene color. The formation of ahaze image model is as following:
( ) ( ) ( ) (1 ( ))I x J x t x A t x= + (1)
Where I is the observed luminance and is also the inputhaze image, J is the scene radiance and is also the restored
haze-free image, A is the global atmospheric light which can
be obtained by using dark channel prior, and tis the mediumtransmission describing the portion of the light that is not
scattered and reaches the camera. Theoretically, the goal of
haze removal is to recoverJ,A, and tfrom I.
The dark channel prior proposed by Hes approach [7] is a
kind of statistics of the haze-free outdoor images. It is based
on a key observation that most local patched in the haze-freeoutdoor images contain some pixels which have very low
This work is sponsored by National Natural Science Foundation of Chinagrant DUE 90820302
978-1-4244-3709-2/10/$25.00 2010 IEEE
-
7/28/2019 HazeREmoval_SingleImage_Luminance
2/4
intensities in at least one color channel. Formally, for an
imageJ, we define
{ , , } ( )( ) min ( min ( ( )))dark c
c r g b y xJ x J y
= (2)
Where cJ is a color channel ofJand ( )x is a local patch
centered at x. we call
dark
J the dark channel ofJ, and we alsocall the above statistical observation the dark channel prior.Using this prior with the haze imaging model, the value of air
light A can be directly estimated to recover a high quality
haze-free image. In our approach, we also use the haze imagemodel and dark channel prior to remove the haze effect.
B. Retinex used in luminance componentRetinex theory deals with compensation for illumination
effects in images. The primary goal is to decompose a given
image S into two different images, the reflectance image R,
and the illumination image L, such that, at each point (x, y) inthe image domain, S(x, y) = R(x, y) *L(x, y). The Retinex
methodology was motivated by Land's landmark research of
the human visual system [8]. The benefits of such
decomposition include the possibility of removing
illumination effects, enhancing image edge, and correcting the
colors in images by removing illumination induced color
shifts.
The original idea of our approach lies in the process of
obtaining transmission map. According to Hes approach, the
transmission map is in fact an alpha map with clear edge
outline and depth layer of the scene objects, which is used toestimate the thickness of the haze. While the MSR algorithm
covers the features of multi-scale by adjusting the scale
parameters, the algorithm synthesize the advantages of the
dynamic range compression, edge detail enhancement of the
small scale, and the balance of the color of the medium and
large scale. So image sharpness, dynamic range compressionof gray level, contrast enhancement and color balance can be
realized at the same time. Applying MSR to the luminance
component of the haze image in YCbCr color space, then does
linear transformations and median filter operation, the
transmission map of our approach can be obtained, so it
realizes the automatic and quick acquisition of transmission
map.
III. IMAGE HAZING REMOVAL METHOD BASED ONLUMINANCE COMPONENT
A. Estimating the Transmission Based On luminanceComponent
Note that, the haze imaging Equation (1) has a similar
form with the image matting equation. A transmission map is
exactly an alpha map. Therefore, Hes approach applied a soft
matting algorithm [9] to refine the transmission. While the
specific steps of our approach to estimate transmission map
based on luminance component is as following:Here, we first choose the transformed color space. For a
RGB image, in order to separate its luminance component
from chroma component, color space should be transformed to
HIS, YUV, etc to assure the steps can be done only in
luminance component. Recently, the most widely used is HIS.However, the transformation between HIS and RGB has to do
trigonometric function calculation, which costs much time and
needs a lot of calculating works. So we choose YCbCr, the
reason is that the transformation between YCbCr and RGB is
refer to simple algebraic operation, so it needs little
calculating work and the speed of the calculation is relativelyfast.
Then, we use MSR algorithm to the luminance component,
the process can be expressed as follows:
1
( , ) (log ( , ) log[ ( , )* ( , )])N
m n n
n
R x y w Y x y F x y Y x y=
= (3)
Where Rm(x, y) is the output of the transformation to the
luminance component image by using MSR algorithm, N is
the number of the scale (usually choose three), Wn is the
weight corresponding to each scale, Y(x, y) is the distribution
of luminance image, Fn(x, y) is the nth wrap around functionwhose form is Gaussian, and the around function should cover
all kinds of scale. Experiment shows that the scale should be
chosen with small, medium and large value for most images.
The measurement of choosing scale is keeping the color of
luminance component image balance and making the contrastof the component image enhance. The sum of the scale value
should be one. The new luminance image obtained by this step
is the fundamental of obtaining transmission map whose
function is similar to the one obtained by soft matting
algorithm in Hes approach.Note that the differences between the new luminance
image and the alpha map obtained by soft matting algorithm,
we want to make both image tend to the same effect andfunction. The solution to this problem is using the
transmission adjusting parameter C minus the value of each
pixel. Specifically, there are two steps to adjust transmission
map: the first one is determining the value of atmospheric
light A. We first picking the top 0.1% brightest pixels in the
dark channel, the pixel with highest luminance in the input
imageIis selected as the atmospheric light. The second step is
determining the value ofCon condition that the value ofA is
determined. The value ofCis application-based. We fix it to
1.08 for all results reported in this paper.
B. Medium Filter and Recovering the Scene RadianceAccording to the haze image model we can recover thescene radiance with the transmission map. Because the direct
attenuation term J(x)t(x) can be very close to zero, so the
transmission t(x) is restricted to a lower bound t0, a typical
value of t0 is 0.1. The final scene radiance J(x) is recovered
by:
0
( )( )
max( ( ), )
I x AJ x A
t x t
= + (4)
-
7/28/2019 HazeREmoval_SingleImage_Luminance
3/4
From the above expression we can see that the nearer the
outline of scene objects in the new luminance image to the
input haze image, the easier to lose details in restored image.We can see from Fig. 1, if the transmission map shown as the
green continuous curve is similar to the input image shown as
the black continuous curve, the result will be a flat curve
shown as the red line after division operation. On the contrary,
if the green curve is more flat, the red curve will be more
variable, which means the dehaze result will contain moredetails. Therefore, in order to flatten the outline of scene
objects in luminance image, and make the haze removal image
contains more details, we use medium filter to the luminance
image. Fig. 2 shows an example of the transmission map
obtained without or with medium filter and their respective
restored haze-free result. Besides, experiment results show
that if the local patch chosen too small, the scene objects in
haze removal image seem lack of three-dimensional
appearance; while if the local patch is too large, the color of
scene objects in restored image seems too dark. When the
patch size is set to 5*5 for a 200*200 image, the haze removal
effect is the best. In this way, the function of the new
luminance image after medium filter can estimate the
thickness of the haze.
Since now we already know the input haze image I(x),
the transmission map t(x) and the atmospheric lightA, we can
take these values into (4) to obtain the final restored haze-freeimageJ(x). Fig. 3 (b) is the transmission map estimated from
color image Fig. 3 (a) using our approach. Fig. 3 (c) is our
final recovered scene radiance to Fig. 3 (a).
Figure 1. The input haze image is shown as the black continuous curve. The
transmission map is shown as green continuous curve. The restored haze-freeresult is shown as the red line.
(a) (b) (c) (d)
Figure 2. Medium filter result. (a) restored haze-free result without filter. (b)
transmission map obtained without filter. (c) restored haze-free result withfilter. (d) transmission map obtained with filter.
(a) (b) (c)
Figure 3. Haze removal result. (a) input color image. (b) transmission map
after using method based on luminance component. (c) final dehazing image
using our method.
IV. EXPERIMENTAL RESULTS AND ANALYSISA. Qualitative Comparison
In our experiments, we perform the improved method by
executing MATLAB on a PC with 3.00GHz Intel Pentium
Dual-Core Processor. Fig. 4 shows a comparison betweenresults obtained by Fattal [5] and our algorithm. It can be seen
that our approach outperforms Fattals in situations of dense
haze. Fattals method is based on statistics and requires
sufficient color information and variance. If the haze is dense,
the color is faint and variance is not high enough for his
method to reliably estimate the transmission. Next, we
compare our approach with Tans work [6] in Fig. 5. The
colors of his result are often over saturated. Since his
algorithm is not physically based and may underestimate the
transmission. Our method can generate comparable results
without any halo artifacts. We also compare our method with
Hes very recently work [7] in Fig. 6. The overall result of our
approach is approximately the same as Hes approach.
Compared with the input haze image, the contrast and clarity
of the images are enhanced significantly. Fig. 7 allows the
comparison of our results with three state of the art visibility
restoration algorithms: Fattal [5], Tan [6] and He [7]. Noticethat the results obtained with our algorithm seems visually
close to the result obtained by He, with less halo artifacts
compared with Tan.
Figure 4. Comparison with Fattals work[5]. Left: input image. Middle:
Fattals result. Right: our result.
Figure 5. Comparison with Tans work[6]. Left: input image. Middle: Tans
result. Right: our result.
Figure 6. Comparison with Hes work[7]. Left: input image. Middle: Hes
result. Right: our result.
Figure 7. From left to right, the input image and the results obtained by
Fattal [5], Tan [6], He [7] and our algorithm.
-
7/28/2019 HazeREmoval_SingleImage_Luminance
4/4
B. Quantitative EvaluationTo quantitatively assess and rate these four methods, we
first consider the processing time. Our method takes about 10-12 seconds to process a 600*400 pixel image in Matlab
programming environment. From Tab. I, we can see that our
method is about two times and three times faster than the
speed of Hes and Fattals method respectively and is far
faster than Tans method.
Then, we use the method of visible edges segmentation
proposed in [10]. Here, we transform the color level image to
the gray level image, and use three indicators e , r and tocompare two gray level images: the input image and the
restored image. The visible edges in the image before and after
restoration are selected by a 5% contrast thresholding, which
allows computing the rate e of edges newly visible afterrestoration. Then, the ration r of the average gradient afterand before restoration is computed. This indicator r estimatesthe average visibility enhancement obtained by the restoration
algorithm. At last, the percentage of pixels which becomes
completely black after restoration is computed.
These indicators e , r and are evaluated for Fattal [5],Tan [6], He [7] and our approach on four images, see Tab. II,
Tab. III and IV. In each approach, the aim is to increase the
contrast without losing some visual information. Hence, good
results are described by high values of e and r and lowvalues of . From Tab. II, we deduce that depending of the
image, Tan algorithm has more visible edges than He, Fattaland our algorithm. From Tab. III, we can order the four
algorithms in decreasing order with respect to average
gradient: Tan, our, He and Fattal. This confirms our
observations on Fig. 4, Fig. 5, Fig. 6 and Fig. 7, and one can
notice that the contrast of the Tans algorithm has been
increased probably too strongly. Tab. IV gives the percentage
of pixels which become completely black after restoration.
Compared to others, ours and Hes algorithm give the smallest
percentage.
TABLE I. COMPARISON OF ALGORITHMS RUNNING TIME (S)Figure Fattal Tan He Our
Fig. 4(441*450)
Fig. 5(835*557)
Fig. 6(1000*327)Fig. 7(576*768)
29.8756
67.9213
47.235464.9538
324.4260
746.7254
249.2972527.5627
15.8185
42.3780
23.153932.1535
8.4680
29.7660
12.609025.5160
TABLE II. RATE e OF NEW VISIBLE EDGESe Fattal Tan He OurFig. 4
Fig. 5Fig. 6
Fig. 7
1.01
7.210.94
0.23
1.54
7.622.74
1.71
1.42
7.361.22
0.80
1.47
7.400.92
1.78
TABLE III. RATIO r OF THE AVERAGE GRADIENTr Fattal Tan He Our
Fig. 4Fig. 5
Fig. 6
Fig. 7
1.292.18
1.26
1.09
1.592.32
1.48
1.40
1.342.24
1.28
1.27
1.402.30
1.38
1.37
TABLE IV. PERCENTAGE OF PIXELS WHICH BECOMES COMPLETELY
BLACK AFTER RESTORATION(%)
Fattal Tan He OurFig. 4
Fig. 5
Fig. 6Fig. 7
1.11
1.02
0.881.05
1.29
0.49
0.360.76
0.82
0.34
0.170.30
0.92
0.27
0.250.29
V. CONCLUSIONSIn this paper, we have proposed an automatic image hazing
removal method based on luminance component, for singleimage haze removal. The approach introduces the MSRalgorithm to the luminance component with the color space
transform, and then does subtraction operation and medianfilter on the component. Its main advantages are its speed andno user interaction is needed. Results shows that it achieves asgood or even better results compared to the main present-dayalgorithms as illustrated in the experiments. The proposedmethod may be used with advantages as pre-processing inmany systems, such as surveillance, topographical survey,intelligent vehicles, etc.
REFERENCES
[1] Ming-Jung Seow, Vijayan K. Asari. Ratio rule and homomorphic filterfor enhancement of digital colour image[J]. Neurocomputing, 2006,69(7):954-958
[2] Fabrizio Russo, An Image Enhancement Technique CombiningSharpening and Noise Reduction[J], IEEE Transactions onInstrumentation and Measurement, 2002, 51(4):824-828
[3] Dongbin Xu, Chuangbai Xiao. Color-preserving defog method for foggyor haze scenes [C]// Proceedings of the 4th International conference oncomputer vision theory and applications. IEEE Press, 2009: 69-73
[4] McCartney E J. Optics of Atmosphere: Scattering by Molecules andParticles[M]. New York: John Wiley and Sons, 1976:23-32.
[5] R. Fattal. Single image dehazing. In ACM SIGGRAPH, 2008:1-9[6] R. Tan. Visibility in bad weather from a single image. CVPR, 2008:1-8[7] Kaiming He, Jian Sun, and Xiaoou Tang. Single Image Haze Removal
using Dark Channel Prior. IEEE Conference on Computer Vision andPattern Recognition (CVPR), June 2009
[8] E. H. Land, The Retinex Theory of Color Vision, Sci. Amer., Vol. 237,pages 108-128, 1977.
[9]
A.Levin, D.Lischinski, and Y.Weiss. A closed form solution to naturalimage matting. CVPR, 1:61-68,2006.
[10]N. Hautiere, J.-p. Tarel, D. Aubert, and E. Dumont. Blind contrastenhancement assessment by gradient ratioing at visible edges. ImageAnalysis & Stereology Journal, 27(2):87-95, June 2008